How do you run OpenAI models on a PC without restrictions?
Download OpenAI's open-weight models. gpt-oss-20b runs within 16GB of memory on a normal PC; gpt-oss-120b needs a single 80GB GPU. Both are Apache 2.0 licensed. Install Ollama, run `ollama pull gpt-oss:20b`, and you get no rate limits, no subscription, no cloud — but the model's own safety training still applies.
Why — the first-principles explanation
"Without restrictions" bundles four different limits together, and running locally removes three of them completely while barely touching the fourth. Sorting them out is the whole answer.
Rate limits and cost vanish. When the model runs on your machine, nobody is metering your tokens. Ask it 10,000 questions at 3am; the only bill is electricity. Privacy limits vanish too — your prompts never leave the building, which is the real reason many businesses run models locally. And connectivity vanishes: it works on a plane, in a basement, during an outage.
The fourth limit is different. The model's behavior is baked into the weights. Refusals aren't a filter bolted on top that you can remove by unplugging the internet — they're a pattern learned during training, distributed across billions of numbers. Downloading the weights gets you the same trained behavior, offline. People do publish community fine-tunes that sand that training down, but note what that means: someone spent GPU hours retraining. It's surgery on the model, not a settings toggle. And it changes nothing about the law. A local model doesn't create legal cover for anything you'd be liable for otherwise.
The practical setup is genuinely easy now. OpenAI released gpt-oss-120b and gpt-oss-20b under the Apache 2.0 license — permissive enough for commercial use, no copyleft, no patent trap. The 20b model has 21 billion parameters but activates only 3.6 billion per token, a Mixture-of-Experts trick that's exactly why it squeezes into 16GB. The 120b (117B total, 5.1B active) wants a single 80GB GPU like an H100.
One wrinkle worth knowing: these models were trained with OpenAI's harmony response format, and they misbehave without it. Use Ollama, LM Studio, vLLM or llama.cpp — they handle it for you.
An example that makes it click
It's the difference between renting a car and owning one. Renting means a meter, a contract, a return date, and the rental company knows every place you drove. Buying one outright means you drive as much as you want, at any hour, and nobody logs your trips.
But here's the part people miss: the car you bought still has the speed governor the factory installed. Owning it didn't remove that. You could take it to a shop and have the governor cut out — that's a fine-tune, real work by someone with tools. And either way, owning the car has never once made a speeding ticket go away.
How to do it
- Check your hardware. gpt-oss-20b needs about 16GB of memory (a 16GB+ GPU, or an Apple Silicon Mac with 16GB+ unified memory). gpt-oss-120b needs a single 80GB GPU such as an H100 or AMD MI300X.
- Install Ollama from ollama.com — it's the shortest path on Windows, macOS and Linux and handles the harmony response format for you.
- Pull the model: run `ollama pull gpt-oss:20b` (or `ollama pull gpt-oss:120b` if you have the hardware).
- Run it: `ollama run gpt-oss:20b`. You now have a local chat with no rate limits, no subscription and no network calls.
- Prefer a GUI? Use LM Studio instead — `lms get openai/gpt-oss-20b` — which gives you a chat window and model management without the terminal.
- Serving it to an app? Use vLLM or Ollama's built-in OpenAI-compatible endpoint, so existing OpenAI API client code points at localhost with almost no changes.
- Unplug the ethernet and test. If it still answers, you've confirmed nothing is going to the cloud.
Key facts
- OpenAI's gpt-oss-120b (117B total parameters, 5.1B active per token) runs on a single 80GB GPU such as an H100 or AMD MI300X.
- OpenAI's gpt-oss-20b (21B total parameters, 3.6B active per token) runs within 16GB of memory.
- Both gpt-oss models are released under the Apache 2.0 license — no copyleft restrictions and no patent risk, permitting commercial use and modification.
- Both models use MXFP4 quantization of their Mixture-of-Experts weights, which is what makes the memory footprints achievable.
- Ollama commands: `ollama pull gpt-oss:20b` and `ollama run gpt-oss:20b`; LM Studio: `lms get openai/gpt-oss-20b`.
- The gpt-oss models were trained on OpenAI's harmony response format and require it to function correctly; Ollama, LM Studio, vLLM and llama.cpp handle this automatically.
▶ The 60-second explainer (script)
How do you run OpenAI models on a PC without restrictions? First, untangle what 'restrictions' means — because it's four different things, and going local kills three of them outright. One: rate limits. Gone. Nobody's metering tokens on your own machine. Ask it ten thousand questions at 3am; the bill is electricity. Two: cost. Gone. No subscription. Three: privacy. Gone — your prompts never leave the building. That's the actual reason most businesses do this. Bonus, it works offline. On a plane. In an outage. But the fourth one is different, and this is where the honest answer lives. The model's behavior is baked into the weights. Refusals aren't a filter bolted on top that disappears when you unplug the internet — they're patterns learned in training, spread across billions of numbers. Downloading the weights gets you the same trained model, offline. Community fine-tunes that sand that down exist, but that's surgery — someone burned GPU hours. Not a toggle. And it doesn't change the law one bit. Now the easy part. OpenAI released gpt-oss-120b and gpt-oss-20b under Apache 2.0. The 20b has twenty-one billion parameters but only activates three point six per token — that's why it fits in sixteen gigs. Install Ollama, type ollama pull gpt-oss twenty-b, and you're running.
What authoritative sources say
People also ask
Can I run ChatGPT itself offline?
No. GPT-5-class models are closed and cloud-only. What you can run locally are OpenAI's open-weight gpt-oss models, which are capable but not the same models behind ChatGPT.
Does running locally remove the model's refusals?
No. Safety behavior is trained into the weights, so it comes with the download. Removing it requires actually fine-tuning the model — and it doesn't change your legal responsibility for what you do.
What hardware do I need?
gpt-oss-20b fits within 16GB of memory — a 16GB+ GPU or an Apple Silicon Mac with 16GB+ unified memory. gpt-oss-120b wants a single 80GB GPU like an H100.
Is it legal to run these commercially?
The Apache 2.0 license permits commercial use and modification without copyleft restrictions. OpenAI also publishes a gpt-oss usage policy alongside the models.
Will it be as good as ChatGPT?
Not quite. gpt-oss-120b was reported near-parity with OpenAI's o4-mini on core reasoning benchmarks — strong, but below the frontier cloud models.